Current AI Speeds Up Its Push to Build Open AI for Everyone

The $400 million nonprofit wants AI to work the way the early web did. Six months of shipping suggest the idea has outgrown the manifesto stage.

Current AI, the Paris-born nonprofit trying to turn artificial intelligence into public infrastructure, has compressed a year's worth of launches into the first half of 2026. Since February the organization has unveiled an offline AI device with the Indian government, deployed its first $3.2 million in grants, released a no-login open-source chatbot during the AI for Good Summit in Geneva and signed a development pact with Sakana AI, Japan's most valuable startup.

The premise behind the sprint is blunt. Every frontier AI system belongs to a private company, and Current AI chief executive Ayah Bdeir argues that a technology reshaping daily life needs a counterpart the public can own. Her reference point is deliberate. AI should be "like the World Wide Web, available to anyone, for free," she said in an interview with TechCrunch published Sunday.

The Paris Pledge Behind the Pipeline

Current AI launched in February 2025 at the AI Action Summit in Paris, founded by Martin Tisné with seed money of roughly $100 million from the French government. President Emmanuel Macron used the summit stage to invite additional backers into the partnership, and endorsements followed from UN Secretary-General António Guterres and European Commission President Ursula von der Leyen. Commitments have since passed $400 million, with the Ford Foundation, the MacArthur Foundation, Google DeepMind and Salesforce among the named funders. Reports have also connected the Patrick J. McGovern Foundation and Schmidt Sciences to the effort. The stated ambition is $2.5 billion mobilized over five years, and Bdeir has stressed that the money arrives as philanthropy, from funders rather than investors.

Bdeir took the CEO role in January.

She came from Mozilla, where she led AI strategy and advised the organization's leadership on its reinvention for the AI era. Before that she founded littleBits, the modular electronics company that put snap-together circuit kits in front of millions of schoolchildren before selling to Sphero in 2019. She trained at the American University of Beirut and the MIT Media Lab, and her hardware instincts showed up almost immediately in what Current AI chose to build first.

Suno Sutra: An AI for India That Never Touches the Cloud

The organization's first collaborative build appeared on February 20 at the IndiaAI Impact Summit in New Delhi. Working with the Digital India Bhashini Division, the language technology arm of India's Ministry of Electronics and IT, Current AI produced Suno Sutra, Hindi for "listening chronicles." The handheld unit describes its surroundings in text or audio across all 22 of India's official languages, and it does the work entirely on the device. No connectivity, no cloud, no account, no data leaving the user's hands.

Three models run inside it at once. A vision model interprets what the camera captures. Bhashini's speech recognition and translation systems handle the language layer, while a text-to-speech engine reads answers back aloud. The hardware carries a camera, a screen, a microphone and a speaker, and the full design is open source, published for developer communities to adapt.

Bdeir's stated motivation is a representation gap: India holds hundreds of languages and dialects, she told TechCrunch, and today's AI systems reflect hardly any of them.

In June the partners escalated. Bhashini and Current AI, joined by Kalpa Impact, opened the VYOMA Innovation Challenge, a competition for open-source, voice-first AI that functions without internet access. Twenty shortlisted teams will receive developer kits built on the Suno Sutra platform, with prizes worth up to ₹80 lakh, just under $100,000, and deployment opportunities inside central and state government departments. Target use cases span governance, education, agriculture and healthcare.

Where the First $3.2 Million Landed

On June 19 Current AI announced its pilot grant cohort, four organizations selected around one theme: cultural preservation.

Masakhane, the grassroots research network focused on natural language processing for African languages, took the largest share at $2.25 million. The Kenya-anchored community works across more than 50 languages spoken by over a billion people, and the grant backs both its institutional independence and a push into domain-specific datasets. Health and agriculture data come first. Education follows.

Lebanon's Institute for Worldmaking received $285,000 to convert Arab cultural history and contemporary practice into public, machine-readable knowledge. The funded plan includes two AI-ready cultural databases and an auditable, retrieval-based research agent, wrapped in governance rules that let Arab communities decide how their knowledge enters AI systems at all.

Portal sem Porteiras, which works with Indigenous communities in the Brazilian Amazon and the Cerrado, also received $285,000. Its tools are offline-first and locally hosted, so recordings and trained models stay inside the territory, and the communities themselves decide which tools get built.

The fourth grant, $290,000, went to the African Internet Rights Alliance in Kenya, which is developing reusable audit tools so regulators and civil society groups can scrutinize AI systems that already make consequential decisions across the continent.

The cohort is engineered to share what it learns, so an audit method developed in Nairobi or a consent framework tested in the Amazon becomes usable by every other grantee.

Consent as a Design Requirement

Bdeir draws a hard line between this work and Big Tech's multilingual expansion, which she characterizes as market growth pursued with little regard for consent or context. Her sharpest example, offered to TechCrunch, concerns Indigenous languages in which missionary Bible translations become training data before the communities involved have set any rules. Linguists estimate that around half of the world's spoken languages are endangered, and Bdeir argues that English-dominated models accelerate the loss, because a system unable to speak a language cannot carry the culture encoded inside it.

On data ownership her position leaves little room for interpretation: whatever governance model wins out in a given community, the default should never be a Silicon Valley company enriching a small circle of shareholders. None of the four grantees has fully solved community data governance yet, and Current AI treats that openness as the point. Each has built the question into the work instead of letting complexity become a reason to hand the decision to a corporation or a state.

Alpha Chat and the Potluck Playbook

July brought the most public-facing releases so far, timed around the AI for Good Summit in Geneva. They followed a strategy refresh published June 30, which recast the nonprofit as a convener for a fragmented public-interest AI field.

On July 1 came the Open Source AI Gap Map, an evaluation of 24,626 open-source AI projects intended to show which layers of the stack already exist and which are missing. The map builds on earlier work by open-source AI researchers, including groups convened at Columbia University and teams at Hugging Face.

Eight days later Current AI shipped Alpha Chat, a free chat interface that requires no login and runs on web and mobile. A coalition of ten organizations assembled it in seven weeks; Hugging Face and Mozilla took part, with the MIT Media Lab among the other contributors. One partner supplied a base model trained on consensually gathered data through a fully open pipeline. Others brought safety tooling or compute. Current AI says its own addition was orchestration, the coordination layer that made separate components behave like a single product.

The chatbot is framed as an argument as much as a product. In its launch post, the organization called Alpha Chat evidence that "a fully open source, sovereign, community-configured AI is already possible."

The release sits inside a wider campaign the nonprofit calls the AI Potluck, an explicit rejection of the race metaphor that dominates the industry. The pitch asks aligned organizations to each contribute their strongest component, whether a foundation model, fine-tuning capability, benchmarks, safety tooling or compute. The resilience logic is the dinner table's: if one dish fails to arrive, the meal survives. A proprietary stack, by contrast, can strand millions of users on a single company's policy change.

The Sakana Pact and the Sovereignty Test

The partnership map now stretches to Tokyo. Current AI signed a memorandum of understanding with Sakana AI in March, and the two plan a shared open-source stack serving Japanese language and culture alongside communities across the Global South that mainstream systems overlook.

Sakana is an unusual counterpart for a public-interest nonprofit. Founded in 2023 by former Google researchers David Ha and Llion Jones together with Ren Ito, the startup closed a $135 million Series B last November at a valuation near $2.65 billion, making it Japan's most valuable startup. Its research bet runs against the industry's compute maximalism: rather than training enormous models from scratch, Sakana evolves and merges existing open models, an approach it promotes under the banner of sovereign AI.

The overlap with Current AI is philosophical as much as technical. Both organizations treat cultural and linguistic fit as a first-order engineering requirement rather than a localization afterthought, and both contend that capable AI does not require the largest possible model.

The Scale Question

The obvious critique writes itself. A $3.2 million grant round is a rounding error next to the capital flooding into frontier labs, and a seven-week chatbot will not top any benchmark chart.

Bdeir rejects the yardstick, telling TechCrunch that scale is Big Tech's paradigm rather than hers. The measure she offers instead is transmission: an Indigenous elder in the Amazon using a tool built in Kenya to pass ecological knowledge to the next generation in their own language. Current AI's own strategy makes a matching concession on capability, arguing the open stack does not need to beat frontier labs on every benchmark, only to get close enough that on price and on control it wins for the uses that matter.

There is now a deadline attached to the rhetoric. Current AI has committed to a working prototype of a fully integrated public-interest AI stack by June 2027, an assistant running on open, sovereign infrastructure with provenance and governance designed in from day one. Success metrics are being drafted with partners, including targets for how many startups and public agencies adopt the stack as their primary infrastructure.

For a movement long on manifestos, the deadline supplies the thing it has rarely had: a shipping date.